Unsupervised Anomaly Segmentation at High Resolution with Patch-Divide-and-Conquer and Self-Ensembling

Abstract

In modern industries, quality assurance is becoming an increasingly relevant topic, as is the detection of anomalies. As this is a very difficult task, AI methods are usually utilized. However, this leads to a problem: it is very difficult to obtain data in this area, especially anomalous data. Therefore, in recent years, more and more research has been carried out in the field of unsupervised segmentation of anomalies. However, most of the current approaches work on low resolution images. Therefore, we have developed Patch-DC (Patch-Divide-and-Conquer), a technique that can be combined with existing techniques to improve their performance, especially on high resolution images. We also present another enhancement technique, SE (Self-Ensembling), which can be utilized with existing techniques to enhance performance. To test our techniques, we have created the BP (Bearings and Plates) dataset, with high resolution images for anomaly segmentation tasks. The techniques are evaluated on this dataset, as well as on the commonly used MVTec dataset [ 6 ], and outperform state-of-the-art techniques in many categories. The dataset and the code can be found at https://www.kaggle.com/datasets/hendrikmeininger/bp-bearings-and-plates-dataset and https://github.com/HendrikMeininger/AnomalySegmentation respectively.